I tried 1.6.5 and I crashed that too. I was using the interaction term generating code in the manual to generate interaction terms for each of the 12 months for a number of variables, so I ran the interaction term code 12 times, then did an arch regression with the outputted lists of interaction terms and a bunch of other dummy variables and a lagged Y term, then used fcast to forecast over about 170 observations. the dataset is about 19000 observations large. I put an exit in after the interaction term generating code and gretl crashed there, but it was also crashing before while just running the arch regression.
If you still have the script from my last arch error report, it's remarkably similar to that, but with a bigger dataset and without the squares and cubed terms.
Chris